The report
An OpenAI research report dated November 4 describes a mid-October automated audit of 5,034 Linux code targets, producing 2,308 candidate findings, 1,724 dynamically reproduced vulnerabilities and 948 distinct local privilege-escalation exploit results in controlled settings. It separately reports 291 information-disclosure proofs of concept. These are stages and effect types of a security research pipeline, not a count of real compromised servers or fixed Linux releases.
The full model API-credit cost was $1,213,430, roughly $1,279 per reported LPE. That quotient excludes all human verification, infrastructure and downstream maintainer costs. An earlier comparison run on GPT-5.6 Sol yielded 166 reported distinct LPEs for $513,690, under a related but not identical research period. Neither total is a price list for publicly exploitable vulnerabilities.
The validation failure is as instructive as the scale. In an initial set of 200 purported root-escalation successes, reviewers found 20 that exploited a different zero-day from the one the task assigned. The pipeline added root-cause checks. It also encountered a separate QEMU TCG vulnerability inside its test machinery. A root shell can therefore be real yet misattributed to the wrong component.
At the end of the ten-day scan and manual review, OpenAI reported 61 upstream fixes, 43 staging patches and 299 findings under public discussion. Those figures are dated workflow states, not 403 shipped fixes. In a separate hardened-target kernelCTF-style experiment, the pipeline obtained 25 disclosure PoCs and 21 LPEs, but disabled the SLAB_VIRTUAL mitigation for compatibility; the results were not accepted contest bounty awards. The canonical Linux case preserves the funnel, attribution limitation and patch uncertainty, while the exploit-factory field note examines why executable effects and accepted fixes are different outputs.